面向机器人地形分类的大视野多光谱光场相机机理研究

Terrain classification is one of the key technology in environmental perception for off-road mobile robots. The increasing demand for higher perception capability from robots calls for a better means of information acquiring with larger degree of depth and wider scope. Traditional multi-sensor data fusion based methods have lots of practical problems considering the constraints of the cost, volume, power and system complexity, demanding a new sensor capable of acquiring high-dimension information with large field of view. It is the first time that this proposal induces the light field into the terrain classification for robots, with a novel design of multi-spectrum light field camera with large field of view. The new camera extends the dimension of the acquired data from 2D to 4D, with an extra extension in spectrum dimension as well. A new generalized array imaging model is proposed to enlarge the scope of applicable mirror and simplify the calibration, making the dense 3D reconstruction possible. To improve the accuracy of reconstruction, a new method for wide field stereo matching and dense reconstruction is presented. Finally the resulting multi-spectrum 3D information with large field of view is applied to terrain classification task, providing a more effective solution for passive and low power consumption sensor based robot terrain classification.

地形分类是现代越野机器人环境感知的关键技术之一。随着机器人对地形分类要求的不断提高，对环境信息采集的深度和广度都提出了更高的要求。传统的采用多种主被动式传感器通过数据融合提高性能的方法，在实际应用中面临着系统成本、体积、功耗和复杂度等诸多方面的现实问题。对具有大视野高维信息采集能力的新型传感器的需求十分迫切。本项目首次将光场采集引入到机器人地形分类中，提出了大视野多光谱光场相机的设计方案，将传统相机从2维扩展到4维，同时实现了光谱维度的扩展。提出了该相机的统一阵列模型，扩大了镜面适用范围，简化了标定并使得致密重构成为可能；提出了大视野球面立体匹配和致密重构方法，提高了重构的精度；最后将获得的大视野多谱深度信息应用于机器人的地形分类中，为基于被动式低功耗传感器的地形分类提供了一套有效的解决方案。

地形分类是现代越野机器人环境感知的关键技术之一。传统的采用多种主被动式传感器通过数据融合提高性能的方法，在实际应用中面临着系统成本、体积、功耗和复杂度等诸多方面的现实问题。对具有大视野高维信息采集能力的新型传感器的需求十分迫切。本项目首次将光场采集引入到机器人地形分类中，提出了大视野多光谱光场相机的设计方案，将传统相机从二维扩展到四维，同时实现了光谱维度的扩展。提出了该相机的统一阵列模型，扩大了镜面适用范围，简化了标定并使得致密重构成为可能；提出了大视野球面匹配和多立体致密重构方法，提高了重构的精度；提出了基于深度神经网络的融合多谱与深度信息的语义分割方法，并应用于机器人的地形分类中。项目的成功实施，为提升越野机器人的地形分类能力提供了一整套从传感器到算法的完整解决方案。